About This Article
This article was created using an automated generation workflow leveraging generative AI. It compares cases where the output format is fixed versus not fixed in the instructions to the LLM, using the same dummy meeting minutes.
Verification Status: 📘 General LLM usage principles confirmed, specific model hardware testing not yet conducted
When making requests to LLMs, deciding not only the content but also the output format in advance makes it easier for humans to review.
First, Request the Same Text in Two Ways
9月15日 会議 田中: 見積書を水曜までに更新する。 佐藤: 顧客へ木曜に確認する。 次回会議は金曜10時。
First, we ask as follows: この議事録を整理してください。 Next, we do the following.
この議事録を整理してください。 次の見出しを必ずこの順で出してください。 - 決定事項 - 担当者と期限 - 次回予定 - 不明点 書かれていない内容は推測せず「記載なし」としてください。
Check Here
Instead of focusing on writing quality, we check whether four specific items—person in charge, deadline, next schedule, and points of uncertainty—can be verified every time. In free-form text, even if it is easy to read, deadlines and unclear points can sometimes get buried in the text.
Why Format Specification Works
Because LLMs generate natural language flexibly, simply saying "organize this" does not uniquely determine what should be prioritized. Specifying output items allows humans to create a checklist for evaluating the generated content. However, format specification does not guarantee factual accuracy, so cross-checking against the original text is still necessary.
Change Just One Place
Finally, we add only 不明点には、確認が必要な理由も1行で付けてください。. We observe whether this makes it easier to transition to the next action, even if the output becomes longer.
For Business Use
In meeting minutes, inquiry organization, incident reports, and specification checks, embedding items that humans need to review later into the output format makes re-verification much easier. Comply with organizational AI usage rules, use only information permitted for input, and cross-check generated results against the source text.
Conclusion
A good prompt is not a long incantation. By comparing before and after fixing the output items with the same input, improvements can be observed rather than just felt.
